Created by: LeoZhao-Intel

I met a issue related with mkldnn when I tried to enable transpose mkldnn op in ParallelExecutor for BERT training.

throw exception in TransposeMKLDNNHandler:: AcquireTranspose()

PADDLE_ENFORCE((transpose_p != nullptr) || (is_reusing_ == false),
                   "Fail to find convolution primitive in device context");

For current implementation in Compute(), a hash key is used for create TransposeMKLDNNHandler, and this key is generated by

const std::string key = platform::TransposeMKLDNNHandler::GetHash(
        nchw_tz, axis, ctx.op().Output(framework::GradVarName("X")));

This method is ok for common case in Executor or NativeExecutor since the only instance for one op, and even for multi-instances inference case, SetMkldnnThreadID() is designed for identifying op. But for ParallelExecutor case, it is specific, it duplicates program execution by user defined instance number, it looks like you have multiple GPU cards in one machine. Let's come to the root cause for my case, for ParallelExecutor on CPU, no place to SetMkldnnThreadID() to identify different mkldnn op instances on same program, simply speaking, the key generation method is not enough anymore. For my current temporary solution, I add kernel instance pointer 'this' into hash key, so it can be simplified with just 'this', and then it works. I think it is not a issue just on Transpose, it may be a generic issue for all mkldnn ops on ParallelExecutor.